Multi-label classification [1] is a generalization of multiclass classification, which is the singlelabel problem in categorizing instances with only one or two classes. Multi-label problems exist in several domains, such as document classification [2],[3], text categorization [4],[5], social network [6], music emotions categorization [7],[8]. Previous research on multi-label text classification has involved traditional machine learning algorithms such as k-Nearest Neighbours [9],[10], Naive Bayes [11],[12], Support Vector Machine [13],[14], Logistic Regression [15]. In addition, compared to the traditional algorithm mentioned, it has certain limitations in terms of large-scale dataset training [17].

Just like other traditional single-label classifications, multi-label classifications have limitations when data labels are small [18]. In this case, a large-scale dataset based on previous research is used [19] to overcome multi-label classification in news articles. While news articles consist of several long sentences and can change their meaning if there are missing sentences. Therefore, Recurrent Neural Network (RNN) was chosen to solve this problem, since the recurrent structure is very suitable for long variable text processing [20]. One of deep learning methods proposed in this study is RNN architecture by applying the Long Short-Term Memory (LSTM) which able to expands the memory [21]. However, during training, traditional RNN has gradient vanishing and exploding problem, which can be solved by LSTM [22]. LSTM has different processing with a common RNN model. Another difference is an additional signal given from a time-step to the next time-step, which called context or memory cell.